Diversity analysis

This is an attempt at a streamlined and yet complete, relatively a priori/non- snooped model analysis. Some parts of the analysis have been moved into separate .R files: the overall workflow is

(mmd_procdata is a little bit out of date: Enric has processing machinery to extract lat/long of centroid of largest polygon (?) (x/y in the data set) and area …)

source("mmd_utils.R")

Packages used/versions:

load_all_pkgs()
theme_set(theme_bw())
library('pander')  ## for tables
sapply(pkgList,function(x) format(packageVersion(x)))

##         lme4        bbmle        broom  broom.mixed      lattice 
##     "1.1.15"     "1.0.20"      "0.4.3"      "0.0.1"    "0.20.35" 
##    gridExtra      ggplot2     ggstance        dplyr        tidyr 
##        "2.3" "2.2.1.9000"        "0.3"      "0.7.4"      "0.7.2" 
##       tibble        remef       r2glmm 
##      "1.4.2" "1.0.6.9000"      "0.1.2"

Data from Enric (includes area and lat/long coordinates):

L <- load("ecoreg.RData")

Limit all observations with all predictor variables > 0 and no biome 98/99; set biome as factor. Log and scale/center variables as appropriate. This leaves us with 29 variables and 620 observations.

Model fitting

This can now be done semi-automatically (i.e., fit all combinations of random effects) by using the function fit_all from the mmd_utils.R file (sourced above), e.g. fit_all(response="mbirds_log")

However, the mmd_fitbatch.R code has already run all random-effects model combinations for all four response variables, so we can extract just one set of models (and profiles) this way:

L <- load("allfits.RData")
names(allfits)

## [1] "plants_log" "mamph_log"  "mbirds_log" "mmamm_log"

focal <- "mamph_log"
results <- allfits[[focal]]
## profiles (from best model for each response variable/set of fits only)
L <- load("allprofs.RData")
names(profList) <- names(allfits)
pp <- profList[[focal]] 

subset(aa2 <- AICtabx(results),dAIC<10)

##                                           dAIC df singular
## biome=int/flor_realms=int/biome_FR=diag    0.0 20    FALSE
## biome=diag/flor_realms=int/biome_FR=int    3.0 20     TRUE
## biome=diag/flor_realms=int/biome_FR=diag   3.9 24     TRUE
## biome=diag/flor_realms=diag/biome_FR=int   5.3 24     TRUE
## biome=int/flor_realms=diag/biome_FR=diag   5.5 24     TRUE
## biome=int/flor_realms=int/biome_FR=full    7.2 30    FALSE
## biome=diag/flor_realms=diag/biome_FR=diag  8.7 28     TRUE

## get the name of the best model from all of the different RE models
best_names <- sapply(allfits,get_best_name)
## create a list of the best model for each taxon
best_models <- Map(function(x,n) x[[n]],
                   allfits,best_names)
## e.g. ecoreg$birdres <- residuals(best_models[["mbirds_log"] ])
best_df <- sapply(best_models, function(x) attr(logLik(x), "df"))

To get residuals use residuals(best_model,re.form=NULL); this uses all of the random effects.

Best non-singular models (by AIC):

pander(data.frame(taxon=names(allfits),best_model=best_names,
                  df=best_df),row.names=FALSE)

taxon	best_model	df
plants_log	biome=int/flor_realms=int/biome_FR=diag	20
mamph_log	biome=int/flor_realms=int/biome_FR=diag	20
mbirds_log	biome=int/flor_realms=int/biome_FR=diag	20
mmamm_log	biome=int/flor_realms=diag/biome_FR=diag	24

Diagnostics?

best_model <- best_models[[focal]]
grid.arrange(plot(best_model,type=c("p","smooth")),
             lattice::qqmath(best_model),nrow=1)

rr <- ranef(best_model)

The best arrangement of the random effect plots will vary somewhat depending on which components have single vs multiple terms … the biome_FR plot will be tall in any case (because there are many levels), but the widest plots will be those with diag or full specifications rather than intercept-only models.

ranefplots <- lattice::dotplot(rr)
do.call(grid.arrange,c(ranefplots[2:3],list(nrow=1)))

reorder_var <- "NPP_cv_sc"
if (!is.null(reorder_var) && reorder_var %in% names(rr$biome_FR)) {
    ## only try to reorder if the variable is actually there ...
    rr$biome_FR <- rr$biome_FR[order(rr$biome_FR[[reorder_var]]),]
}
rr3 <- as.data.frame(rr$biome_FR) %>%
    tibble::rownames_to_column("br") %>%
    ## mutate(br=reorder(br,order(NPP_cv_ctr))) %>%
    gather(key=term,value=value,-br)
ggplot(rr3,aes(value,br))+facet_wrap(~term,nrow=1)+geom_point()+
    theme(panel.spacing=grid::unit(0,"lines"))+
    labs(x="",y="")

VarCorr(best_model)

##  Groups      Name        Std.Dev.  
##  biome_FR    Feat_cv_sc  1.6995e-01
##  biome_FR.1  NPP_cv_sc   7.2985e-02
##  biome_FR.2  Feat_log    6.4446e-02
##  biome_FR.3  NPP_log     7.1105e-05
##  biome_FR.4  (Intercept) 2.0458e-01
##  biome       (Intercept) 3.2121e-01
##  flor_realms (Intercept) 2.6413e-01
##  Residual                4.3587e-01

coefficient plot of all models tried

Check difference between Wald and likelihood profile confidence intervals. In principle profile CIs are more accurate - if the computations have run reliably … but I would probably be conservative and take the wider of the two.

FIXME: Wald CIs only give fixed-effect confidence intervals. Could combine, or sort out problems with profiles (or go back to glmmTMB/brms/etc. …)

profci <- confint(pp,parm="beta_")
waldci <- confint(best_model,method="Wald",parm="beta_")

Pick CI type:

ci_method <- "Wald" ## or 'profile' or 'boot'

get_coeftab <- function(model,ci_method="Wald") {
    cc <- confint(model,method=ci_method)
    coeftab <- tidy(model) %>% filter(term !="(Intercept)")
    ## drop spurious numeric values
    tt <- gsub("\\.[0-9]","",coeftab$term)
    ## group separator -> bar
    tt <- gsub("\\.(flor_realms$|biome$|biome_FR$)","|\\1",tt)
    coeftab$term <- tt
    ## nppcv_var <- coeftab$nppcv_var <- grepl("NPP_cv",tt)
    ## re_var <- grepl("^(sd|cor)",tt)  ## & !nppcv_var
    ## otherfix_var <- !re_var & !nppcv_var
    ## fix_var <- !re_var
    ee <- coeftab$estimate
    ## order by magnitude within
    ## (1) random effects
    ## (2) fixed effects
    ## (3) NPPcv
    ## coeftab$term <- factor(tt,
    ##    levels=c((tt[re_var])[order(ee[re_var])],
    ##      (tt[fix_var])[order(ee[fix_var])]
    ##      (tt[nppcv_var])[order(ee[nppcv_var])]))
    mm <- match(tt,rownames(cc))
    mm[is.na(mm)] <- match("sigma",rownames(cc))
    coeftab$lwr <- cc[mm,"2.5 %"]
    coeftab$upr <- cc[mm,"97.5 %"]
    ## coeftab$sd_var <- grepl("^sd",coeftab$term)
    ## coeftab$nppcv_fac <- factor(coeftab$nppcv_var,
    ## labels=c("non-NPPcv terms", "NPPcv terms"))
    return(coeftab)
}
coeftabs <- bind_rows(lapply(best_models,get_coeftab),
                      .id="taxon")
## reorder parameters
mean_est <- coeftabs %>% group_by(term,effect) %>%
    summarise(est=mean(estimate))
fix_var <- mean_est$effect=="fixed"
flevs <- with(mean_est,
              c((term[!fix_var])[order(est[!fix_var])],
                (term[fix_var] )[order(est[fix_var])]))
coeftabs$term <- factor(coeftabs$term,levels=flevs)

ctplot1 <- ggplot(coeftabs,aes(estimate,term,
                     colour=effect))+
    scale_colour_manual(values=c("black","red"),guide=FALSE)+
    ## use point + linerange because some RE are missing std.errors
    ## (could also set NA std.errors to zero)
    geom_point()+
    geom_linerangeh(aes(xmin=lwr, # estimate-1.96*std.error,
                         xmax=upr  # estimate+1.96*std.error
                         ))+
    geom_vline(xintercept=0,lty=2)+
    facet_wrap(~taxon)+  ## ncol=1 might be nice, but too skinny
    labs(x="",y="")
print(ctplot1)

Zoom in a bit: exclude random effects and NPP_log (which is large and significant for all taxa). Abuse warning: coloring significant effects.

ctx <- coeftabs %>%
    filter(effect=="fixed" & term!="NPP_log") %>%
    mutate(effect=factor((lwr*upr)>0))
## suppress message about replacing 'colour'
suppressMessages(print(ctplot1 %+% ctx +
        scale_colour_manual(values=c("black","orange"),guide=FALSE)))

Definitely starting to get into a snooping mindset here. There are not a lot of effects other than NPP_log that are consistently large and significant; perhaps main effects of fire for birds and mammals. Amphibians have considerably larger effects (the last three: Feat cv and its interactions).

plotfun() takes arguments:

• model: fitted model
• xvar (“NPP_log”): x-variable
• auxvar (“Feat_cv_sv”): auxiliary variable (e.g. for examining interactions)
• respvar (equal to model response by default): response variable
• aux_quantiles: (0.1, 0.5, 0.9) quantiles of auxiliary variable to predict
• pred_lower_lim (-3) : lower cut off values (log scale)
• data (ecoreg)
• re.form (NA) which RE to include in predictions (default is none)

ggplot1 <- plotfun(best_model)
print(ggplot1)

FIXME: (1) ? relabel axes with absolute rather than log values ?

fix_NAs <- function(rem,model) {
    if (!is.null(nastuff <- attr(model.frame(model),"na.action"))) {
        return(napredict(nastuff,rem))
    } else return(rem)
}
rem1 <- fix_NAs(remef(best_model,ran="all"),best_model)
if (length(rem1)==nrow(ecoreg)) {
    ## if it worked ...
    ecoreg$rem1 <- rem1
    ## update previous plot:
    plotfun(best_model,respvar="rem1")
}

ecoreg$rem2 <- fix_NAs(remef(best_model,ran="all",
                             fix=c("NPP_log","NPP_cv_sc","Feat_cv_sc"),grouping=TRUE),
                       best_model)
ecoreg$rem3 <- fix_NAs(remef(best_model,ran="all",
                             fix=c("NPP_log","NPP_cv_sc","Feat_log"),grouping=TRUE),
                       best_model)
## FIXME: drop dimensions on Feat_cv_sc upstream
gg1 <- ggplot(ecoreg,aes(c(Feat_cv_sc),rem3))+
    geom_point(aes(colour=biome,shape=flor_realms))+
    geom_smooth(group=1)+
    theme(legend.box="horizontal")
print(gg1)

\(R^2\) across taxa

Only a few effects have partial \(R^2\) values of more than a few percent: NPP (of course), fire (for mammals and ?birds?), and fire CV (for amphibians). Everything else is going to be pretty subtle (provided of course that we trust this particular way of estimating \(R^2\)).

all_rsq <- bind_rows(lapply(best_models,r2beta),.id="taxon") %>%
    mutate(Effect=reorder(Effect,Rsq))
rsqplot <- ggplot(all_rsq,aes(Rsq,Effect,colour=taxon))+
    geom_pointrangeh(aes(xmin=lower.CL,xmax=upper.CL),
                   position=position_dodgev(height=0.5))+
    scale_colour_brewer(palette="Dark2")+
    scale_x_log10(limits=c(1e-2,1),oob=scales::squish)+
    labs(y="")
print(rsqplot)

Models without biome/realm interaction

• explore degree of among-biome/among-realm variability in effects
• look at random effects to see which biomes/realms have high/low (or unusual) effects

L <- load("allfits_restr.RData")
as.data.frame.coef.mer <- function(x,...) {
    class(x) <- "list"
    lapply(x,tibble::rownames_to_column,var="grp") %>%
        bind_rows(.id="grpvar") %>%
        gather(term,condval,-c(grpvar,grp))
}
ranefs_restr <- bind_rows(lapply(allfits_restr,
                 function(x) as.data.frame(ranef(x))),.id="taxon")
coefs_restr <- bind_rows(lapply(allfits_restr,
                 function(x) as.data.frame(coef(x))),.id="taxon")

cd1 <- filter(coefs_restr,grpvar=="biome",term != "log(area_km2)")
gg_restr_biome <- ggplot(cd1,
                         aes(condval,grp,colour=taxon,shape=taxon))+
    geom_point()+facet_wrap(~term,scale="free_x")+
    scale_colour_brewer(palette="Dark2")+theme(legend.pos="bottom")

print(gg_restr_biome)

• better ordering? I think it’s currently ordered by plant intercept term (which makes sense: trop/subtrop moist at the top, mangrove at the bottom …)
• leave out plants entirely?

print(gg_restr_biome %+%
      filter(coefs_restr,grpvar=="flor_realms",term != "log(area_km2)"))

needs work: looking at predicted values rather than deviations from population mean, the values for different taxa are very far apart, need to look at them separately for each taxon … ??

vars_restr <- bind_rows(lapply(allfits_restr,
               function(x) as.data.frame(VarCorr(x))),.id="taxon") %>%
    mutate(grp=gsub("\\.[0-9]*$","",grp)) %>%
    filter(grp!="Residual", !is.na(var1))
ggplot(vars_restr,aes(sdcor,var1,colour=taxon,shape=taxon))+
    geom_point()+facet_wrap(~grp)+
    scale_colour_brewer(palette="Dark2")+
    theme(legend.pos="bottom")+
    labs(y="",x="standard deviation of random effect")

spatial fits

So far …

library(gamm4)
## add spatial smooth to fixed effect formula
spatform <- update(formula(best_models[[1]],fixed.only=TRUE),
                   . ~ . + s(y,x,bs="sos",
                             m=-1 ## play with smoothing order
                             ))
## extract random effect formula and drop LHS
ranform1 <- formula(best_models[[1]],random.only=TRUE)[-2]
spat1 <- gamm4(spatform,random=ranform1,
               data=na.omit(ecoreg))
class(spat1) <- "gamm4"  ## why list()?

?smooth.construct.sos.smooth.spec mgcv:::plot.sos.smooth

par(mfrow=c(1,2))
plot(spat1$gam,theta=0,phi=120,scheme=0,pch=1)
plot(spat1$gam,scheme=1,pch=1)

plot(y~x,data=ecoreg)

tidy.gamm4 <- function(x,...) {
    r <- tidy(x$mer,...)
    r$term <- gsub("^X","",r$term)
    return(r)
}
spatcoefs <- dplyr::bind_rows(lapply(list(orig=best_models[[1]],
                             spat=spat1),tidy,effect="fixed"),
                 .id="model") %>%
    filter(term!="(Intercept)") %>%
    mutate(term=reorder(factor(term),estimate))
ggplot(spatcoefs,aes(estimate,term,
                     xmin=estimate-1.96*std.error,
                     xmax=estimate+1.96*std.error,
                     colour=model))+
    geom_pointrangeh(position=position_dodgev(height=0.5))+
    geom_vline(xintercept=0,lty=2)+
    scale_colour_brewer(palette="Set1")

n.b.: for simple models need random=~(1|biome), random=~1|biome fails with “model frame and formula mismatch in model.matrix()”

To do

• further examine models without biome/realm interaction?
• examine models to see when profile CIs can be used; fix profiles?
• spatial models
  • automate/run for all cases
  • extract residuals, re-check autocorr
  • other info/plotting methods for smooth component? plotting arguments? (see ?smooth.construct.sos.smooth.spec)
  • is extracting the mer component and doing inference etc. OK?
  • should this be done from the start?

To do (fancy)

• More on Bates et al. approach?
• Try with Julia??
• Factor-analytic approach (e.g. Jeon and Rockwood (2017))?

Jeon, Minjeong, and Nicholas Rockwood. 2017. PLmixed: Estimate (Generalized) Linear Mixed Models with Factor Structures. https://CRAN.R-project.org/package=PLmixed.